Model-Based Online Learning and Adaptive Control For A Human-Wearable Soft Robot'' Integrated System

Article
The International Journal of

Robotics Research
Model-based online learning and adaptive 1–21
Ó The Author(s) 2019
control for a ‘‘human-wearable soft robot’’ Article reuse guidelines:

sagepub.com/journals-permissions
DOI: 10.1177/0278364919873379
integrated system journals.sagepub.com/home/ijr
Zhi Qiang Tang1 , Ho Lam Heung1, Kai Yu Tong1 and Zheng Li2
Abstract
Soft robots are considered intrinsically safe with regard to human–robot interaction. This has motivated the development
and investigation of soft medical robots, such as soft robotic gloves for stroke rehabilitation. However, the output force of
conventional purely soft actuators is usually limited. This restricts their application in stroke rehabilitation, which requires a
large force and bidirectional movement. In addition, accurate control of soft actuators is difficult owing to the nonlinearity of
purely soft actuators. In this study, a soft robotic glove is designed based on a soft-elastic composite actuator (SECA) that
integrates an elastic torque compensating layer to increase the output force as well as achieving bidirectional movement.
Such a hybrid design also significantly reduces the degree of nonlinearity compared with a purely soft actuator. A model-
based online learning and adaptive control algorithm is proposed for the wearable soft robotic glove, taking its interaction
environment into account, namely, the human hand/finger. The designed hybrid controller enables the soft robotic glove to
adapt to different hand conditions for reference tracking. Experimental results show that satisfactory tracking performance
can be achieved on both healthy subjects and stroke subjects (with the tracking root mean square error (RMSE) \0.05 rad).
Meanwhile, the controller can output an actuator–finger model for each individual subject (with the learning error RMSE
\0.06 rad), which provides information on the condition of the finger and, thus, has further potential clinical application.
Keywords
Model-based learning control, soft robot, human–robot interaction
1. Introduction upper (Balasubramanian et al., 2008) and lower (Park

et al., 2014) limbs. In particular, wearable soft robotic
Soft robots have attracted much attention in recent years gloves (In et al., 2015; Polygerinos et al., 2015) have been
owing to their distinctive features compared with rigid designed for hand rehabilitation. Such soft robotic gloves
robots. Soft robots have found applications in a variety of are different from the soft robotic hands (Catalano et al.,
fields, including astronautics (Villoslada et al., 2018), deep 2014; Deimel and Brock, 2016) that are used as prostheses.
sea exploration (Galloway et al., 2016), and medicine The output force of soft robotic gloves is usually limited,
(Awad et al., 2017; Huang et al., 2018). Among medical
applications, soft robots for rehabilitation are increasingly 1
being employed owing to their inherent safety, light weight, Department of Biomedical Engineering, The Chinese University of Hong
Kong, Shatin, Hong Kong SAR
and increased potential for portability and efficacy, in con- 2
Department of Surgery and the Chow Yuk Ho Technology Centre for
trast to conventional rigid rehabilitation devices (Dovat Innovative Medicine, The Chinese University of Hong Kong, Shatin,
et al., 2008; Hu et al., 2013), which have limitations of Hong Kong SAR
heavy weight and large size that may hinder upper-limb
Corresponding authors:
movement and performance of activities of daily living
Kai Yu Tong, Department of Biomedical Engineering, The Chinese
(ADLs). Robot-assisted hand rehabilitation for stroke sur- University of Hong Kong, Shatin, Hong Kong SAR.
vivors has demonstrated improved motor recovery after Email: [email protected]
rehabilitation training (Balasubramanian et al., 2010;
Takahashi et al., 2007) and more clinicians are interested to Zheng Li, Department of Surgery and the Chow Yuk Ho Technology
Centre for Innovative Medicine, The Chinese University of Hong Kong,
use robotic devices in clinical rehabilitation. Up to now, Shatin, Hong Kong SAR.
soft robots have been designed for assisting movements of Email: [email protected]
2 The International Journal of Robotics Research 00(0)
Fig. 1. 3D-printed soft robotic glove.

Fig. 2. Exploded view of one 3D-printed soft-elastic composite
whereas soft robotic hands need to be able to supply a actuator (SECA).
strong force when grasping various objects in their role as
replacements for missing hands. Soft robotic gloves need
to interact with human hands for assisting hand movement, humans is an essential consideration with regard to the
and therefore hand conditions are important factors in their control of soft robots for rehabilitation. Existing controllers
design and control, whereas these conditions are not rele- proposed for the manipulation of soft robots can be cate-
vant for prosthetic robotic hands. The development of con- gorized into three types (George Thuruthel et al., 2018):
trol algorithms remains a challenge for soft robots, owing model-free, model-based, and hybrid controllers. Model-
to the nonlinearity of soft actuators and their interaction free controllers are usually based on machine learning tech-
with the environment. niques or empirical methods. Such controllers have great
Pioneering soft robotic gloves (In et al., 2015; advantages in highly nonlinear, non-uniform, and unstruc-
Polygerinos et al., 2015) mainly provided flexional move- tured environment situations where modeling is almost
ment to grasp objects, but could not produce sufficient impossible. Although Lyapunov stability and convergence
extensional force for stroke patients who are unable to can be proved in some model-free adaptive controllers
extend their fingers because of spasticity. This led to a situ- (Parks, 1966), analytical methods for the stability and con-
ation in which a stroke survivor was able to grasp an object vergence analysis of popular machine-learning-based
but could not release it. Therefore, this type of system model-free controllers nowadays are difficult to derive
could not assist the training of stroke patients with spasti- owing to probabilistic or non-parametric properties
city. Consequently, another type of fabric-based bidirec- (Braganza et al., 2007; Kocijan, 2016; You et al., 2017). In
tional soft robotic glove was designed to offer active applications where stability is the first priority and system
assistance with both finger flexion and extension (Yap behavior needs to be predictable, model-based approaches
et al., 2017). However, the overall size of this bidirectional are preferred. Model-based controllers (Best et al., 2016;
actuator was twice that of the earlier type (around 30 mm Morrow et al., 2016; Tang et al., 2019) usually need analy-
in thickness), making the system more cumbersome to tical models to derive the controller. These controllers have
wear. more accurate and reliable performance than model-free
In the present study, a new soft robotic glove is designed controllers for uniform soft manipulators in known envir-
to assist both flexional and extensional movement while onments. However, they usually require well-defined dyna-
retaining a small size (15 mm in thickness) by incorporat- mical models for the soft robots, which may not be easy to
ing an elastic torque-compensating layer into the soft actua- construct on the basis of rigid-body assumptions (Trimmer,
tor, as shown in Figures 1 and 2. This design can provide 2014). Advanced accurate modeling methods (Renda et al.,
flexional torque on pressurization, and the compensating 2018; Zhang et al., 2016) have limited use owing to their
layer can provide extensional torque on depressurization. computational complexity and the cost of the sensors
Such a hybrid design also exhibits a significantly reduced required. Hybrid controllers (Braganza et al., 2007;
degree of nonlinearity compared with a purely soft actua- Reinhart et al., 2017) combine model-free and model-based
tor, with consequent benefit to dynamic modeling of the controllers, and are usually based on an analytical model to
actuator. capture the main part of the system’s intrinsic properties
The interaction of soft robots with the environment is a and a data learning model to compensate dynamic uncer-
critical challenge faced by soft robotic modeling and con- tainties. Hybrid controllers are a viable and promising
trol. In particular, the interaction between soft robots and approach for soft robot control (George Thuruthel et al.,
Tang et al. 3
2018). In this study, we design a hybrid controller for our repetitive stretching exercise training for stroke patients. In
system. addition, an adjustable reference trajectory is designed,
It is important to consider the human finger as the inter- which takes both individual range of motion differences
action environment for a wearable soft robotic glove. For and human–robot interaction comfort into consideration.
stroke rehabilitation, finger spasticity is the main difficulty To the best of the authors’ knowledge, this article is the
that needs to be overcome (Yue et al., 2017). Spasticity first to take finger spasticity into account in the design of a
results from impaired reflex function and induces a number controller for a soft robotic glove. The contributions of this
of changes in muscle properties, such as stiffness, fibrosis, article include the following.
and atrophy (Thibaut et al., 2013). It has been shown that
joint stiffness can be treated as a constant when the finger Simplification (i.e., linearization) of the control for soft
joint moves at a fixed velocity (Kamper et al., 2006, 2002). robots by incorporating elastic materials, using the
Each individual has a constant joint stiffness, but stiffness SECA as the example.
varies from person to person. Hence, it is necessary to Consideration of the interaction environment (finger)
design an effective control algorithm that allows a soft in the development of the dynamic model and control
robotic glove to adapt to differences in finger spasticity. problem of the soft actuator–finger system.
Our proposed control algorithm is a novel integration of Development of a model-based online learning adap-
model predictive control (MPC) and iterative learning con- tive control algorithm for the soft actuator–finger sys-
trol (ILC) that simultaneously achieves model learning and tem with theoretical proof and experimental validation.
reference trajectory tracking of targets. In contrast, previous A proposal for use of the learned model when investi-
MPC and ILC combinations have aimed at improving gating stroke finger conditions.
tracking performance only. Cueli and Bordons (2008) and
Lee et al. (1999) focused on how to incorporate iterative The remainder of this article is organized as follows.
information into an MPC controller and improve its track- The soft robotic glove and an overview of the control sys-
ing performance by redesigning the MPC cost function. tem are described in Section 2, and the model of the soft
Bukkems et al. (2005) and Liu and Xi (2014) emphasized actuator–finger system is presented in Section 3. The com-
ILC controller design as a way of improving ILC tracking plete model learning adaptive control algorithm is
performance. With regard to model learning control, explained in detail in Section 4. Results of simulations are
Ichikawa et al. (2006) and Underwood and Husain (2010) presented in Section 5. Tests of the proposed method on
developed an online parameter identification method based six human subjects are described in Section 6. The poten-
on a recursive least-squares algorithm to control tial clinical application and limitations of the model are dis-
permanent-magnet synchronous motors, and Bouffard et al. cussed in Section 7. Finally, our conclusions from this
(2012) proposed a learning-based model predictive control- investigation are given in Section 8.
ler based on the extended Kalman filter method for control
of a quadrotor. The learning technique in our proposed
method can output an accurate model for the soft actuator–
2. System description
finger system within specified error tolerance. In addition, Our system consists of a soft robotic glove and pneumatic
the learned model parameters can provide insight into the control hardware. Each item is described in a subsection in
behavior of the system, which has clinical potential for the the following.
investigation of finger conditions.
In this article, we propose a model-based online learning
and adaptive control algorithm for a wearable soft robotic
2.1. Soft robotic glove
glove, taking account of its need to interact with stroke sur- We have previously designed a soft robotic glove that
vivors. We describe the complete architecture of the control assists stroke survivors in performing ADLs (Heung et al.,
algorithm as well as a theoretical proof of its capabilities. 2019). In the present study, the soft robotic glove is further
Our hybrid controller utilizes a dynamic model structure modified via our proposed control algorithm to adapt to dif-
for the soft actuator–finger system and a model learning ferent stroke survivors suffering from finger spasticity. Our
method to allow updating of model parameters based on soft robotic glove consists of a hand base that enables the
observed data in real time. The learning technique enables 3D-printed SECA to be attached to a hand with Velcro
the soft robotic glove to adapt to different hand conditions. straps, as shown in Figure 1. Figure 2 shows the structure
Our method provides the reference tracking performance of the SECA, which is composed of metacarpophalangeal
that is necessary for rehabilitation training of stroke survi- (MCP) and proximal interphalangeal (PIP) segments corre-
vors with the aim of improving motor functions (Ao et al., sponding to actuation of the MCP and PIP joints of the
2015; Timmermans et al., 2009). Because hand rehabilita- human finger. When wearing the soft robotic glove, the
tion training is usually a repetitive process (Takahashi et al., position of hand base can be adjusted to align MCP and
2007), the reference trajectory is taken to be a periodic sig- PIP segments of the SECA with MCP and PIP joints of the
nal, and our controller is designed for periodic reference human finger, respectively. Once aligned, the hand base is
tracking. This reference tracking can provide the necessary fixed. Then Velcro straps are used to secure each finger on
Fig. 3. Demonstration of bidirectional movement of the SECA on a wooden hand. (a) Resting state of the wooden index finger with
attached torsional springs (stiffness 2 N/mm for MCP and 0.6 N/mm for PIP). The sum of the MCP and PIP angles in the resting state
is 1108. (b) Extended state of the wooden index finger bound with the SECA at 0 kPa pressure. The sum of the MCP and PIP angles
in the extended state is 608. (c) Flexed state of the wooden index finger bound with the SECA at 87 kPa pressure. The sum of the
MCP and PIP angles in the flexed state is 1058.
the SECA to minimize any gap between the finger and

SECA during the training period. The SECAs for our soft
robotic glove are double-segmented and lack covering on
the distal interphalangeal (DIP) joint because the DIP joint
contributes only 15% to a functional grip (Leibovic and
Bowers, 1994) and stroke patients need to have a sense of
touch at the fingertips when grasping objects (Tong et al.,
2010). Each segment of the SECA consists of a silicone
actuator body, fiber reinforcements, a flex sensor, and a tor-
que compensating layer. On pressurization with fluid, the
fiber reinforcement restricts the radial expansion of each
segment, as a result of which they become elongated in the Fig. 4. Architecture of the pneumatic control setup.
axial direction. The torque compensating layer further con-
strains the bottom elongation, acting as a strain limiting
layer for bending of the MCP and PIP segments. On depres- bending angle. Gerboni et al. (2017) have already shown
surization, the torque compensating layer provides an the high accuracy and repeatability of the selected flex sen-
assisting torque to facilitate effective extension of the MCP sors in instantly measuring bending angles. Transparent
and PIP segments to their initial resting position. Figure 3 fiber reinforcements are embedded in the MCP and PIP
illustrates the flexed and extended states of a wooden index segments.
finger bound to the SECA. Torsional springs with stiffness
2 N/mm for MCP and 0.6 N/mm for PIP are attached to the
MCP and PIP joints of the wooden index finger to simulate 2.2. Pneumatic control hardware
a hand clenched owing to spasticity (Yap et al., 2016). The architecture of the pneumatic control setup is shown in
The silicone actuator body of the double-segmented Figure 4. The control algorithm is implemented on
SECA (length 120 mm, width 18 mm, height 15 mm, and LabVIEW 2015 Version (National Instruments, Corp.,
mass 24 g) is directly 3D-printed from silicone (ACEO USA). The data acquisition device (DAQ; USB-6009,
Silicone GP Shore A 30 from Wacker Chemie AG and National Instruments, Corp., USA) transmits control com-
ACEO, Burghausen, Germany). The selected silicone mate- mands to a proportional solenoid valve (ITV2091-
rial offers high elongation (from 200% to 800%) and hard- 21N2BS5, SMC, Tokyo, Japan) and receives sensor read-
ness (Shore A 30) to prevent failure during bending ings from the solenoid valve and the flex sensors. The air
deformation. An A2 stainless steel plate of thickness 0.2 pump (BTC Diaphragm Pump, Parker Hannifin
mm is adopted as the torque compensating layer. Soft sen- Corporation, OH, USA) supplies air pressure for the
sors (Kim et al., 2018; Soter et al., 2018; Thuruthel et al., SECA. The proportional solenoid valve can regulate the air
2019) are good potential candidates for soft robotic appli- pressure based on command signals from the DAQ. The
cations. Currently, in our system two commercially avail- two flex sensors measure the bending angles of the MCP
able 2:200 (5.6 cm) flex sensors (Flexpoint Sensor System, and PIP segments. The flex sensor data are sampled at
Draper, UT, USA) are positioned between the MCP/PIP 1,000 Hz with 16-bit resolution. The controller frequency
segment and the torque compensating layer to measure the is 10 Hz.
Tang et al. 5
Fig. 5. (a) Illustration of the interaction between SECA and finger. The MCP and PIP segments of the SECA correspond to the MCP
and PIP joints of the finger, respectively. (b) Side view of one segment (MCP or PIP) of the SECA, illustrating the bending state. (c)
Zoomed-in cross-sectional view showing the geometry of the system.
3. System modeling manipulator can be expressed as a function of bending

angle:
This section gives a detailed description of the procedures
for modeling the interaction between the soft actuator and 1 1
the human finger. PEjs = Vj Es e2 = Kjs q2j ð2Þ
2 2
where e is the material strain, qj is the bending angle of
3.1. SECA model segment j, and Kjs is the equivalent stiffness of the silicone
body at segment j.
The fundamental energy problem for soft robots is similar The torque compensating layer here can be viewed as a
to that for rigid robots (Trimmer, 2014). We construct a cantilever beam because one end is fixed in the hand base
dynamic model for each segment of the SECA from an and the other end can bend freely. According to Rao and
energy perspective. The model is based on the assumption Yap (2011), the strain energy of this layer can be expressed
of piecewise-constant curvature (Li and Du, 2013; as
Marchese and Rus, 2016; Webster III and Jones, 2010),
which is widely used in soft robot modeling. Based on the p 4 Et I 2 1 t 2
Lagrangian equations of motion, we first need to find the PEjt = q = Kq ð3Þ
64lj3 j 2 j j
potential and kinetic energy of each segment. Owing to the
light weight of the SECA, we ignore the gravitational where Kjt is the equivalent stiffness of the torque compen-
potential energy. As the fiber reinforcements restrict radial sating layer with respect to segment j, and Et and I are the
expansion, circumferential expansion can also be neglected. Young’s modulus and moment of inertia of the torque com-
The main deformation is elongation of the silicone actuator pensating layer, respectively. Thus, the total potential
body. The potential energy of each segment contains strain energy of each segment is
energy stored in this body and in the torque compensating
layer. For the silicone actuator body, the material volume 1
PEj = PEjs + PEjt = Kja q2j ð4Þ
incurring strain of each segment is 2
where Kja = Kjs + Kjt is the total stiffness of actuator seg-
1 2 1 2
Vj = pr + 2(w + d)r2 pr + 2wr1 × lj ment j.
2 2 2 1 The bending velocities of the silicone actuator body and
ð1Þ torque compensating layer are the same because they are
attached to each other. According to Rao and Yap (2011),
where j 2 fm, pg, with j = m and j = p representing the the kinetic energy of each segment can be expressed as
MCP and PIP segments, respectively, lj is the length of seg-

ment j, r1 and r2 are the inner and outer radii, respectively, 1 3 4 1
w is the chamber gap, and d is the SECA thickness, as KEj = (msj + mtj )lj2 q_ 2j = Jja q_ 2j ð5Þ
2 2 p 2
shown in Figure 5c.
The stressâĂSx-strain relationship of our SECA material where Jja is the equivalent moment of inertia of actuator
can be approximated with a constant Young’s modulus Es segment j, and msj and mtj are the masses of the silicon
when the strain is below 100% (based on data provided by body and the torque compensating layer, respectively, at
Wacker Chemie AG and ACEO). According to Marchese segment j.
et al. (2016), the strain energy of a soft fluidic elastomer The moment generated by pressure is
Z
w
t x_ = Ax + BP
Mja = 2Pr1 x1 + d + dx1
2 y = Cx
Z0 r1 qffiffiffiffiffiffiffiffiffiffiffiffiffiffi
t ð6Þ
+ 2P r12 x22 x2 + w + d + dx2 x = ½ q_ m qm q_ p qp T
0 2 2 3
Mja = Tja P DJmm KJmm 0 0
6 7
6 1 0 0 0 7 ð9Þ
where Tja is the equivalent moment–pressure constant of A=6 6 0
7
Jpp Jpp 7
D K
4 0 5
actuator segment j, t is the thickness of the torque compen-
sating layer, x1 and x2 are integration variables, and P is the 0 0 1 0
h iT
pressure. T
B = TJmm 0 Jpp 0
Then, by using the Lagrangian LGj = KEj PEj , we
can obtain the equation of motion of segment j as follows: C=½0 1 0 1
! where Jj = Jja + Jjf , Dj = Daj + Dfj , Kj = Kja + Kjf , Tj =
d ∂LGj

∂LGj
= Mja Daj q_ j Tja + Tjf , j 2 fm, pg, and y is the total finger angle, which
dt ∂q_ j ∂qj ð7Þ is the sum of the MCP and PIP joint angles. In our physical
) Jja €qj + Daj q_ j + Kja qj = Tja P experiment, the MCP and PIP joint angles qm and qp are
measured by flex sensors, and the pressure P is generated
where Daj is the damping coefficient of actuator segment j, by the proportional solenoid valve.
which accounts for the non-conservative energy. The
dynamic model structure (7) also matches the modeling 4. Control algorithm
results in Marchese et al. (2016).
Based on the model structure described previously, we pro-
pose a model learning adaptive control algorithm to achieve
3.2. Finger joint model our control objectives. Stretching exercise training can help
Widely used finger joint dynamics satisfy the following relax muscles and reduce finger spasticity (Hesse et al.,
second-order linear system with constant coefficients when 2008). Such training is usually applied in the treatment for
the finger joints move at a fixed velocity (Blana et al., stroke patients who are unable to extend fingers owing to
2019; Chen et al., 2014; Kamper et al., 2006, 2002): spasticity and contracture (Yue et al., 2017). Meanwhile,
finger conditions need to be assessed during the training
Jjf €
qj + Dfj q_ j + Kjf qj = Mjf ð8Þ process. The control objectives are to design a stretching
exercise training routine and investigate finger conditions
where Jjf , Dfj , and Kjf are the moment of inertia, damping simultaneously. Technically speaking, we want to achieve
coefficient, and stiffness, respectively, of finger joint j, and satisfactory reference trajectory tracking performance for
Mjf is the external moment exerted on finger joint j. In our stretching exercise training and find the true model para-
application, this external moment is generated by pressure. meters for investigating finger conditions. We give a
Here Mjf = Tjf P, Tjf is the moment–pressure constant of fin- detailed description of our proposed control algorithm in
ger joint j. this section.
3.3. SECA–finger model 4.1. Reference angle design

The incorporation of the elastic compensating layer and the The reference angle is designed for the total finger angle,
finger dynamics enables us to model the soft actuator– which is the summation of PIP and MCP joint angles,
finger system as a linear system. The MCP and PIP seg- because our SECA is an under-actuated system in which
ments of the SECA correspond to the MCP and PIP joints the two fluid chambers of the MCP and PIP segments are
of the human finger, as shown in Figure 5a, and we assume connected. The reference angle is our tracking target.
that the bending angles of the SECA MCP and PIP seg- When choosing the reference angle, we have two consid-
ments are equal to the corresponding finger MCP and PIP erations to bear in mind: (1) the reference angle should
joint angles. This assumption is based on our previous have a smooth trajectory for safe interaction between robot
work (Heung et al., 2019), in which we showed that the dif- and human; (2) it should be adjustable for different ranges
ference between the SECA segment angles and the corre- of motion of the hand. Previous studies (Amirabdollahian
sponding joint angles could be ignored. Hence, the state et al., 2001; Loureiro et al., 2001a,b) have applied mini-
space equation of the SECA–finger system can be written mum jerk theory to design reference trajectories for a reha-
as follows: bilitation robot that helps stroke survivors in upper limb
Tang et al. 7
rehabilitation training. Minimum jerk theory (Hogan, 4.2. Model learning law
1984) is a simple approach to the design of human-like The learning target is to acquire true model parameters. To
smooth trajectories in real-time control situations. Here we achieve this target, a discrete-time transfer function by z-
apply it to design the reference angle. According to mini- transform is first derived for each joint as follows:
mum jerk theory, we need to find a time-dependent func-
tion by minimizing the jerk parameter over the time of u1i z1 + u2i z2
movement: ^qi (k) = Gi Pi (k) = Pi (k) ð14Þ
1 + u3i z1 + u4i z2
Z T 3 2
d r where i is the iteration number, k is the time instance, and
min 3 dt ð10Þ
0 dt ^qi (k) is the model estimated joint angle. Let

T
where t is time and t 2 ½0, T with T the designed time ui = u1i u2i u3i u4i 4×1
period, and r is the time-dependent reference angle func-
ui (k) = ½ Pi (k 1) Pi (k 2) ^qi (k 1) ^qi (k 2) T4 × 1
tion. As a specific form for the reference angle function, a
polynomial is a good candidate because it is convenient for ð15Þ
designers to control the position, velocity, and acceleration
where k = 2, 3, . . . , N , and N = T =Ts, Ts is the sampling
of movement. In the context of minimum jerk theory, a
time. Then we have
fifth-order polynomial has been shown to be sufficient to
create a smooth trajectory (Hogan, 1984). When jerk move-
^qi (k) = ui (k)T ui ð16Þ
ment is minimized, higher-order polynomials will be
reduced to fifth-order polynomials (Amirabdollahian et al., The true model parameters are searched by minimizing the
2002). Hence, we apply a fifth-order polynomial to define error between the actual angle and the model estimated
our reference trajectory: angle, as shown by the following quadratic function:
r(t) = a0 + a1 t + a2 t2 + a3 t3 + a4 t4 + a5 t5 ð11Þ minkEi k22

ui
where ai 2 R, i = 0, 1, , 5 are the polynomial coeffi- s:t: Ei = qi ^qi
cients. Values of these coefficients will affect the position, qî = Fi ui ð17Þ
velocity, and acceleration of the reference trajectory. These
coefficients are determined by a set of boundary condi- qi = ½ qi (2) qi (3) qi (N) T(N 1) × 1
tions. Safety issues and comfortable interaction between Fi = ½ ui (2) ui (3) ui (N ) T(N 1) × 4
the human hand and the soft robotic glove are considered
to set the boundary conditions, listed as follows: where qi is the actual joint angle. Then the Gauss–Newton
method (Mascarenhas, 2014) is applied to update the model
rjt = 0 = r0 , rjt = T = re parameters:
dr dr ð12Þ
jt = 0 = 0, j =0
dt dt t = T ui + 1 = ui + r(FTi Fi )y FTi Ei ð18Þ
d2r d2r where the learning step size r is a constant.

j = 0, j =0
dt2 t = 0 dt2 t = T
where r0 and re are the start and end angles, respectively. 4.3. Reference tracking law
The velocity and acceleration are set to be zero at the begin-
Based on the model learning law, we propose a reference
ning and end of the movement to minimize movement risks
tracking method using iterative learning model predictive
(Amirabdollahian et al., 2002; Hogan, 1984). The reference
control (ILMPC), which is a combination of MPC and ILC.
angle range ½r0 , re will be selected based on stroke patient’s
The actual input pressure consists of the feedback pressure
comfortable range of movement. By minimizing the inte-
ui calculated by MPC and the feedforward pressure vi given
gral (10) and taking account of the constraints (12), we
by ILC.
obtain the following reference trajectory:
First, the updated model provided by the model learning
3 t 4 t 5 law is fed into the internal model predictive controller. The
t
r(t) = r0 + (re r0 ) × 10 15 +6 cost function of the internal MPC is
T T T
ð13Þ hP
+H
min ½Q(r(k) yi (k))2 + Ru2i (k)
ui
where a0 = r0 , a1 = 0, a2 = 0, a3 = 10(re r0 )=T , 3 k =h+1 ð19Þ
xi (k + 1) = Ai xi (k) + Bi ui (k)
a4 = 15(re r0 )=T 4 , and a5 = 6(re r0 )=T 5 .
yi (k) = Ci xi (k)
where Q and R are output and input weighting factors, 4.4. Complete algorithm description
respectively, Q and R are positive scalars, i is the iteration To integrate the modeling learning law and the reference
number, k is the time instance, and h and H are the reced- tracking law into a complete algorithm, we first define the
ing and prediction horizons, respectively. By solving the model learning error and the reference tracking error. The
optimization problem (19), we obtain the feedback pressure criterion for successful model learning is that the model
ui (k) as follows: predicted output should be sufficiently close to the system’s
k actual output, given the same input. Hence, the model
ui (k) = hiêi learning error is defined as the root mean square error
êki = rh T
i xi (k) (RMSE) between the system’s actual output qi and the
model predicted output ^qi :
rh = ½ r(h + 1) r(h + 2) r(h + H) TH × 1
1 T rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
hi = ½ 1 0 TQ
0 1 × H (S iS
i + R
i ) (Q i)
iS 1 2
2 3
i
epi = kq ^qi k2 ð23Þ
R 0 0 N 1 i
60 R 0 7
6 7 Meanwhile, a satisfactory reference tracking perfor-
i = 6 .
R .. .. .. 7
6. 7 mance should be such that the system’s actual output is suf-
4. . . .5
ficiently close to the reference trajectory. Hence, the
0 0 R H ×H
2 3 reference tracking error is defined as the RMSE between
Q 0 0 the reference angle and the actual angle:
60 Q 0 7
6 7 rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
i =6
Q 6 .. .. . . .. 7
7 1
4. . . .5 eri = kr yi k22 ð24Þ
N
0 0 Q H ×H
ð20Þ Finally, we choose the proper error tolerance de to ter-
minate the algorithm. Pseudo-code for our proposed model
2 3
Ci B i 0 0 learning adaptive control is presented in Algorithm 1. The
6 CAB Ci B i 0 7 algorithm convergence analysis is described elaborately in
6 i i i 7
Si = 6

6 .. .. .. .. 7 7 Appendix B.
4 . . . . 5
M1 M2
Ci Ai Bi Ci Ai Bi Ci B i H × H
2 3 Algorithm 1. Model learning adaptive control.
Ci Ai
6 Ci A2 7
6 i 7 1: Initialization i = 0, r 2 ½r0 , re , ui = u0 and vi = 0.
i = 6
T 7 2: Run ILMPC.
6 .. 7
4 . 5 3: do
4: Execute model learning law.
Ci AM
i H ×4 5: Execute reference tracking law.
6: i+ +
Second, based on the idea of ILC (Arimoto et al., 1984), 7: Compute epi and eri .
the feedforward pressure vi is calculated as follows: 8: return ui and yi
9: while jeri eri1 j.de or jepi epi1 j.de
vi + 1 (k) = vi (k) + lei (k + 1)
ei = r yi 5. Simulation results
ð21Þ
r = ½ r(1) r(2) r(N ) TN × 1 To validate the performance of our proposed control algo-
yi = ½ yi (1) yi (2) yi (N ) TN × 1 rithm, we first conduct some simulation experiments using
the LabVIEW 2015 software platform. In this section, the
where the feedforward input gain l is a constant. simulation procedure is presented in detail and the control
The final input pressure is then given by parameters needed to guide the controller design are
discussed.
Pi (k) = ui (k) + vi (k) ð22Þ
where Pi (k), ui (k), and vi (k) are the final input pressure, 5.1. Algorithm simulation procedure
feedback pressure, and feedforward pressure at time
instance k in iteration i, respectively. In our simulation, we assume that the true SECA–finger
system model as
Tang et al. 9
Table 1. Simulation control parameters.
u0 r0 re T Ts Q R H r l dE
m1 08 908 6s 0.1 s 100 1 20 0.4 10 1 × 1025
x_ = Ax + BP angle tracking can be achieved with satisfactory

y = Cx performance.
x = ½ q_ m qm q_ p qp T
2 3 5.2. Parameter sensitivity analysis
5 300 0 0
6 1 0 0 0 7 ð25Þ In our proposed control algorithm, there exist several custo-
6 7
A=6 7 mized parameters, such as learning step size r and feedfor-
4 0 0 10 400 5
ward input gain l. Here we discuss how these parameters
0 0 1 0
affect the performance of our control algorithm. This analy-
T
B=½3 0 5 0 sis will provide some suggestions to help us in designing
C=½0 1 0 1 control parameters.
The model described by (25) is assumed to be unavail-

able for controller design and just serves as the I/O data 5.2.1. Independence of reference trajectory. The time
generator. In actual control situations, the system is exe- period and angular range described in (11) can be adjusted
cuted in discrete time. Here we discretize the continuous- to satisfy actual requirements. Figure 7a shows a reference
angle with time period 4 s and angular range ½208, 1208,
time models in (25) by applying the zero-order hold
that is different from that in Figure 6e. Although in our
method to the inputs with sampling time of 0.1 s. To follow
application the reference angle is designed as (11) for
our control algorithm, we first need to give an initial value
stroke rehabilitation, our proposed control algorithm does
to the model parameter u. To test the efficiency of our algo-
not rely on the reference angle being of this kind. Other
rithm and simplify the choice of initial value, we take the
kinds of reference angle can also be tracked using our con-
initial model parameters for both MCP and PIP joints to be
trol algorithm. Figures 7b and 7c show the simulation
u0 = m1, where m is a constant scalar and
results for tracking a step angle and a trapezoidal angle.
1 = ½1, 1, 1, 1T 2 R4 . The value of m is selected to mini-
For the simulation results shown in Figure 7, with the
mize model learning iterations. From Figure 8b, we can see
exception of the reference angle, all the control parameters
that m = 0:01 is the best with minimum learning iterations
are the same as those in Table 1 and the assumed true
among tested m values. Hence, we adopt m = 0:01 in the
model is also the same as (25). Figure 7 demonstrates that
simulation. The other control parameters for the simulation
our control algorithm can retain good tracking performance
are listed in Table 1.
for various input trajectories. Furthermore, Figure 8a shows
The simulation results are shown in Figure 6. From
that the proposed model learning method can be adaptive
Figures 6b and 6d, we can see that the estimated model
to different types of reference angle including sinusoidal
parameters gradually converge to the true values. After the
wave, step commands, and trapezoid. The convergence of
model learning process, we convert the learned discrete-
model learning is not affected by the choice of reference
time model into continuous-time model by applying the
angle.
zero-order hold method to the inputs. The learned model
parameters and corresponding true model parameters are
listed in Table 2. From the plots (2) and (3) in Figure 6f, we 5.2.2. Effect of initial model value m. Different choices of
see that the error between model estimated angle and actual initial model may influence the model learning convergence
angle converges to 0 after 30 iterations, which demonstrates speed, as shown in the proof of Theorem 1 in Appendix B.
the success of the learning process. Moreover, Figure 6e Here we choose different values of m to see the effects of
shows that the actual total angle (the sum of the MCP and initial model choice. The assumed true model is the same
PIP angles) gradually approaches the reference angle. The as (25). Except for the value of m, all the control parameters
plot (1) in Figure 6f intuitively shows that the reference are the same as those in Table 1. We take the learning pro-
tracking error converges to zero and reaches a steady state cess of the model parameter u4m as a detailed example. u4m is
for iterations greater than 30. one of the MCP model parameters. Figure 8b shows the
The simulation results show that the true model para- effects of different values of m on the model learning pro-
meters can be learned from an initial guess and reference cess. We can see that different initial model values will
Fig. 6. Simulation results. In (a) and (c), the estimated MCP and PIP angles gradually become close to the actual angles. In (b) and
(d), the estimated model parameters of MCP and PIP gradually approach the respective true parameters. In (e), the total angle, given
by the sum of the MCP and PIP actual angles, gradually reaches the reference angle. In (f), the reference tracking error and model
learning errors gradually decrease below error tolerance value.
Table 2. Simulation model parameters.
Km =Jm Dm =Jm Tm =Jm Kp =Jp Dp =Jp Tp =Jp
True model 300 5 3 400 10 5

Learned model 300.21 5.01 3.00 400.29 10.01 5.00
finally converge to the true model parameter, but with dif- learning process of the model parameter u4m as a detailed
ferent convergence speeds. example. Figure 8c shows the impact of different values of
r on the model learning process. We can see that larger val-
5.2.3 Effect of learning step size r. Choosing different ues of r lead to smaller numbers of iterations being
learning step sizes will influence the model learning con- required for convergence when r 2 (0, 1). However, when
vergence speed. We try different values of r to see the r 2 (1, 2), the model parameter learning shows marginal
effects of learning step size. The assumed true model is the stability. When r.2, the model parameter learning will
same as (25). Except for the value of r, all the control para- diverge. This simulation result is again consistent with the
meters are the same as those in Table 1. We also take the theoretical result in Theorem 1 in Appendix B.
Tang et al. 11
(a) (b) (c)

Fig. 7. Different kinds of reference angles. (a) A series of designed reference angles with time period T = 4 s and angular range
½208, 1208, which is different from the reference angle in Figure 6e. (b) A series of 308 step angle commands ranging from 08 to 908,
changing the command at time increments of 2 s. (c) A series of trapezoidal angles with 9 s period ranging from 08 to 908, ramping at
4:58/s linearly.
(a) (b) (c)
(d) (e) (f)

Fig. 8. Impacts of different control parameters on model learning and reference tracking. (a) The model parameter learning is
independent of the choice of reference angle. (b) Different initial model values will finally converge to the true model parameter with
different convergence speeds. (c) When r.2: model parameter learning will diverge; when 1\r\2: model parameter learning will
oscillate around the true value within a bounded region; 0\r\1: the model parameter will converge to the true value, and a larger
value of r will speed up convergence. (d) The reference tracking error will decrease below the error tolerance, and a larger value of l
can help speed up tracking when l satisfies 0\1 lG z\1, whereas l = 0 (1 lG z = 0) and l = 50 (1 lG z = 0:1), which
do not meet the requirement, will cause tracking divergence. (e) and (f) Simultaneously increasing values of r and l in their required
ranges will speed up model learning and reference tracking convergence.
5.2.4 Effect of feedforward input gain l. Different values Equation (25)). This proof is shown in the Theorem 2 in
of the feedforward input gain will influence the reference Appendix B. In particular, a value of l = 0 means that the
tracking convergence speed. We try different values of l to ILC component is not involved in the reference tracking.
see the effects of feedforward input gain. The assumed true These simulation results demonstrate that the ILC compo-
model is the same as (25). Except for the value of l, all the nent is necessary and helpful for our designed control algo-
other control parameters are the same as those in Table 1. rithm. Therefore, an appropriate value of l needs to be
Figure 8d shows the effects of different values of l on the determined in the actual experiments.
reference angle tracking process. We can see that when
l = 10 and l = 20, the tracking error can finally decrease 5.2.5 Effect of varying r and l together. Because r appears
below the error tolerance and that a larger l value will in the discrete-time transfer function and l in the state
speed up tracking of the reference trajectory. However, space equation, it is difficult to express the model learning
when l = 0 and l = 50, the tracking error will diverge or reference tracking error as a function of r and l in an
because 1 lG z = 0 when l = 0 and 1 lG z = 0:1 explicit form. Nonetheless, we can still observe the effects
when l = 50 which do not satisfy the requirement on the control performance when r and l are changed
0\1 lG z\1 (G is the discrete time model of simultaneously. As can be seen from Figures 8c and 8d,
Table 3. Stroke characteristics.
Stroke subjects Age range (years) Stroke onset Gender Stroke type Hemiplegic side ARAT MAS
1 55–59 21 months Male Ischemic Left 9 1

2 50–54 11 months Male Hemorrhagic Right 6 1+
3 50–54 36 months Male Hemorrhagic Left 2 2
1
ARAT (Action Research Arm Test) is an observational measurement for evaluating upper limb function. Full score is 57. A higher ARAT score
indicates better upper limb performance.
2
MAS (Modified Ashworth Scale) is a grading method for spasticity measurement. MAS value ranges from 0 to 4 (0 = No increase in muscle tone; 1
= Slight increase in muscle tone at the end of the range of motion; 1 + = Slight increase in muscle tone throughout the range of motion; 2 = More
marked increase in muscle tone throughout the range of motion). A higher MAS index indicates more severe spasticity. The MAS values here were
measured for finger flexors.
larger values of r and l in their required ranges can indivi- and minimize the risk to subjects (Ao et al., 2015). The
dually speed up learning and tracking convergence, respec- SECA linear elastic bending range and the subject’s com-
tively, and so we explore the effects of increasing r and l fortable range of movement were considered when select-
together. Figure 8e shows the combined effects of r and l ing the reference angle range. All three stroke subjects’
on parameter estimation, and Figure 8f shows the effects on impaired hands could be passively fully extended without
reference tracking. From Figure 8e, we can see that combi- any pain or discomfort. The subjects did not feel any dis-
nations of larger r and l can cause the parameter learning comfort during the experiments. Other control parameters
to converge faster when r and l are in their required ranges. were determined by the simulation results and the actual
A similar observation can be made from Figure 8f. The ref- experimental situations. When the control algorithm started
erence tracking error decreases faster with combinations of running, the training data were continuously collected until
larger r and l. The simulation results in Figures 8e and 8f the termination of the algorithm. The experimental setup is
indicate that combinations of large values of both r and l shown in Figure 9.
can be selected to speed up model learning and reference
tracking convergence under the premise of system stability.
6. Experimental results
The simulation results provide support and guidance for the
following physical experiments. In this section, we describe
some preliminary experiments to demonstrate the feasibility
of our proposed control method.
We used the following protocol to perform physical
experiments. The experiments involved three situations: (1)
the SECA without a human finger (the free bending case);
(2) the SECA with an unimpaired finger (the healthy sub-
ject case); (3) the SECA with a stroke-impaired finger (the
stroke subject case). The experiments were approved by the
Joint Chinese University of Hong Kong–New Territories
East Cluster Clinical Research Ethics Committee (approval
protocol number NCT03286309). Three healthy subjects
(all men, age range 22–24) and three stroke subjects
(detailed information is listed in Table 3) voluntarily parti-
cipated in this pilot study. Each subject gave informed con-
sent before the experiment. During experimentation, each
subject was asked to stay relaxed and the subject’s index Fig, 9. Experimental setup. The control algorithm is
finger (left hand for healthy subjects and impaired hand for implemented on a personal computer (64-bit operating system, i5
CPU and 8 GB RAM) based on LabVIEW 2015 software. The
stroke subjects) was passively moved with the SECA. The
battery supplies electric power for the air pump and proportional
time period of each iteration was taken as 3 s both for flex-
solenoid valve. The air pump supplies air pressure for the SECA
ion and for extension, based on our group’s previous reha- and the proportional solenoid valve regulates air pressure. The
bilitation study (Susanto et al., 2015). The maximum joint data acquisition collects data from the proportional solenoid
torque exerted by SECA was limited to 0.1 Nm with air valve and flex sensors. The PCB board integrates wire
pressure in ½0, 100 kPa and bending velocity was less than connections. The SECA assists finger movement and flex
308 per second, which could ensure experimental safety sensors measure bending angles.
Tang et al. 13
Table 4. Experimental control parameters.
u0 r0 re T Ts Q R H r l d
Free bending m1 108 808 6s 0.1 s 100 1 20 0.02 10 0.01

Healthy subjects ufb 208 808 6s 0.1 s 100 1 20 0.03 15 0.01
Stroke subjects ufb 208 608 6s 0.1 s 100 1 20 0.01 10 0.01
m = 0.002 and ufb are the final learned model parameters of free bending.
Table 5. Experimental learned SECA–finger system model parameters.
Km =Jm Dm =Jm Tm =Jm Kp =Jp Dp =Jp Tp =Jp
Free bending 51.21 5.10 0.25 5.11 3.74 0.03

Healthy subject 1 75.90 7.15 0.37 38.61 4.23 0.31
Healthy subject 2 94.82 8.92 0.46 41.99 5.40 0.32
Healthy subject 3 64.76 5.94 0.32 38.12 3.26 0.34
Stroke subject 1 175.60 14.88 0.75 122.60 10.59 1.18
Stroke subject 2 204.99 17.76 0.97 101.16 10.12 0.81
Stroke subject 3 231.06 21.49 1.05 153.74 13.25 1.49
For free bending, the model parameters are those for the SECA MCP and PIP segments.
6.1. Free bending case 6.2. Healthy subject case

For the free bending situation, the system becomes just the After the free bending experiment, we tested our control
SECA model, and so the free bending experiment can help algorithm on healthy subjects whose hands were unim-
us determine accurate SECA model parameters, thus provid- paired. The free bending results provided an appropriate
ing a basis for the following human subject tests. Similar to initial value for the finger–SECA model. Hence, we took
the simulation procedure, we took u0 = m1 to initialize the the free bending model learning results as the initial model
discrete-time models of the soft actuator’s MCP and PIP seg- for the healthy subject tests. The detailed experimental
ments. Initial model parameter m = 0:002 was empirically results of healthy subject 1 are given in the following and
selected to minimize learning iterations. The control algo- the other two healthy subjects have similar patterns. Before
rithm parameters used in the free bending experiment can be the test, we measured the index finger angle of each healthy
seen in Table 4. The model learning and reference tracking subject’s left hand using a protractor. The finger angle was
results for the free bending situation are shown in Figure 10. around 208 in the resting state. We then adjusted the refer-
The real-time free bending situation is shown in the video in ence angle to lie in the range ½208, 808. The control algo-
Extension 1. We convert the learned discrete-time models rithm parameters used in healthy subjects’ experiment are
into continuous-time models by applying the zero-order listed in Table 4.
hold method to the inputs. The converted continuous-time The experimental results of healthy subject 1 are shown
model parameters are listed in Table 5. in Figure 11. The real-time healthy subject 1 test situation
From Figure 10a we can see that the tracking performance is shown in the video in Extension 2. We also converted
gradually improves and finally satisfies our requirement. the learned discrete-time models into continuous-time
Figure 10b intuitively shows the monotonically decreasing ten- models. The converted continuous-time model parameters
dency of the reference tracking error. Figure 10c shows that are listed in Table 5. From Figure 11 we can see that it
the MCP model estimated angle gradually approaches the takes nine iterations to reach the threshold of conver-
actual MCP angle. Figure 10d also shows that the learning gence. Ten iterations were needed for the other tested two
error of the MCP model gradually decreases below the defined healthy subjects to converge. Figure 11a shows that the
error tolerance value. Meanwhile, we can see that each MCP actual finger angle gradually approaches the reference
model parameter shows a tendency to converge to a certain angle. From Figures 11c and 11e, we can see that the
point, as illustrated in Figure 10g. The learning process of the MCP and PIP model estimated angles gradually approach
PIP model exhibits similar behavior to that of the MCP model. their actual values, respectively. These can be seen more
The PIP model estimated angle gradually covers the actual intuitively in Figure 11d and 11f, where the learning error
PIP angle, and the PIP model parameters show a tendency to of MCP and PIP gradually decreases below the error tol-
converge, as depicted in Figures 10e and 10h. erance value.
Fig. 10. Free bending experimental results. In (a) and (c), the estimated MCP and PIP angles gradually approach their respective
actual values. In (e), the total angle, given by the sum of the actual MCP and PIP angles, gradually reaches the reference angle. In (b),
(d), and (f), the model learning errors and reference tracking error gradually decreases below the error tolerance value. In (g) and (h),
we show the learning processes of the MCP and PIP model parameters.
6.3. Stroke subject case 108. Taking account of the finger spasticity and the sub-
The motor functions of stroke subjects’ paretic upper ject’s comfort, we adjusted the reference angle range to
limbs were assessed by a trained clinical assessor who ½208, 608. The initial model parameters for stroke subjects
was blinded to the experiment. Clinical information on were also based on free bending model learning results.
stroke subjects are listed in Table 3. One stroke subject’s The control algorithm parameters used are listed in Table
(stroke subject 1) detailed experimental results are given 4.
in the following. Before the test started, we measured the The experimental results of the stroke subject test are
MCP and PIP joint angles of the subject’s impaired-hand shown in Figure 12. The real-time stroke subject test situa-
index finger. We found that the subject’s maximum ranges tion is shown in the video in Extension 3. The converted
of voluntary motion for the MCP and PIP joints were both continuous-time model parameters can be seen in Table 5.
Tang et al. 15
Fig. 11. Healthy subject 1 experimental results. In (a) and (c), the estimated MCP and PIP angles gradually approach their respective
actual values. In (e), the total finger angle, given by the sum of the actual MCP and PIP angles, gradually reaches the reference angle.
In (b), (d), and (f), we show the tendency of model learning errors and reference tracking error. In (g) and (h), we show the learning
processes of the MCP and PIP model parameters.
From Figure 12 we can see that it takes 17 iterations to 7. Discussion

reach the threshold of convergence. The other tested two
In this section, we further explore the model learning results
stroke subjects needed 15 and 14 iterations to converge.
in physical experiments and provide a method for deriving
Figure 12a shows that the actual total finger angle gradually
independent finger joint model parameters. We also discuss
approaches the reference angle. From Figures 12c and 12e,
the limitations of the model.
we can see that the MCP and PIP model estimated angles
gradually approach the actual MCP angle, respectively.
From the current experimental results, it takes around 10
7.1. Separate finger joint model parameters
iterations for healthy subjects and 15 iterations for stroke
subjects to converge to the threshold. Stroke subjects need The final learned model parameters in the healthy and
longer iterations than healthy subjects to converge. stroke subjects’ experiments are those of the overall system
Fig. 12. Stroke subject 1 experimental results. In (a) and (c), the estimated MCP and PIP angles gradually approach their respective
actual values. In (e), the total angle, given by the sum of the actual MCP and PIP angles, gradually reaches the reference angle. In (b),
(d), and (f), we show the tendency of model learning errors and reference tracking error. In (g) and (h), we show the learning
processes of the MCP and PIP model parameters.
model, which contains both the SECA and the human fin- Kja Kj Jja
= c1 , = c2 , =1
ger. Here we want to derive independent finger joint model Jja Jj Jjf
parameters. According to Dempster (1955), Peters et al. f
(2002), and Kamper et al. (2002), the index finger length Kj Kja + Kj
) = = c2 ð26Þ
of an adult is around 90 mm and the index finger mass is Jj Jja + Jjf
about 30 g in general. The length of our SECA is 120 mm
Kjf
and its mass is 24 g. The ratio of the moments of inertia ) = 2c2 c1
between the finger and the SECA can be assumed to be a Jjf
constant. Because the average length and mass of an adult’s
index finger are similar to those of our SECA, we assume where c1 can be found from the free bending model learn-
this ratio to be 1, which means that Jja =Jjh = 1. Based on ing results, and c2 from the model learning results for the
this assumption, independent finger joint model parameters healthy or stroke subjects. Here Dfj =Jjf and Tjf =Jjf can be
can be derived as follows: derived in a similar way as in (26). The moment of inertia
Tang et al. 17
Table 6. Finger joint model parameters.
Kmf =Jmf Dfm =Jmf Tmf =Jmf Kpf =Jpf Dfp =Jpf Tpf =Jpf
Healthy subject 1 100.59 9.20 0.49 72.11 4.72 0.59

Healthy subject 2 138.43 12.74 0.67 78.87 7.06 0.61
Healthy subject 3 78.31 6.78 0.39 71.13 2.78 0.65
Stroke subject 1 299.99 24.66 1.25 240.09 17.44 2.33
Stroke subject 2 358.77 30.42 1.69 197.21 16.50 1.59
Stroke subject 3 410.91 37.88 1.85 302.37 22.76 2.95
(Jjf ) is a constant for each subject and, thus, Kjf =Jjf could In our physical experiments, the range of motion that was
be served as a stiffness indicator for each subject. We call tested was restricted to ½08, 908, which lies in the linear
Kjf =Jjf a stiffness indicator as there is no particular physical elastic range of the SECA material. However, if the bend-
meaning for it. Although stiffness values (Kjf ) cannot be ing angle is larger than 908, the SECA tends to be subject
directly obtained in this method, stiffness indicators to nonlinear behavior, for which the current linear assump-
(Kjf =Jjf ) could be used to reflect the stiffness change for tions in the SECA model may no longer be valid. An
each subject. Similarly, Dfj =Jjf and Tjf =Jjf are the damping Euler–Bernoulli beam model with large deflection
indicator and moment–pressure indicator of finger joint j, (Banerjee et al., 2008; Lee, 2002) may need to be applied
respectively. The separate finger joint model parameters are in situations with larger bending angles. In addition,
listed in Table 6. although the assumption that the model coefficients of fin-
From Table 6, we can observe that there exist obvious ger joint dynamics can be treated as constants is valid when
differences between healthy subjects and stroke subjects the finger joints move at a fixed velocity, they will vary if
with respect to the MCP and PIP model parameters. In par- the joint velocities vary. In this case, velocity-dependent
ticular, the stiffness indicator (Kjf =Jjf ) values of stroke sub- coefficients will need to be used in the finger joint
jects are clearly larger than that of healthy subjects. It is dynamics. Assumptions of fixed joint velocity plus con-
reasonable because the spasticity of stroke subjects is obvi- stant curvature limit the current model. However, one of
ously more severe than that of healthy subjects. This com- the main contributions in this study would be the concept
parison is also consistent with the results of Kamper et al. and framework of the iterative learning approach to assess
(2006, 2002). Meanwhile, the stiffness indicator values the environment that the soft robot interacts. Variable velo-
show a similar tendency with the spasticity levels measured city and non-constant curvature models could be further
by modified Ashworth scale (Ansari et al., 2008) in explored in the same control framework.
Table 3. Such similar tendency indicates that a larger stiff-
ness indicator value may imply more severe finger
spasticity. A larger sample size is needed to provide more 8. Conclusion
evidence for this implication. The stiffness indicator can
provide guidance for hand rehabilitation training and may We have proposed a model-based online learning adaptive
be used to quantify MCP and PIP joint spasticity condi- control algorithm for a wearable soft robotic glove, taking
tions. In addition, stiffness indicator values as well as other its interaction with its environment, i.e., the human finger,
model parameters can help improve our control algorithm. into account. The asymptotic stability of the control algo-
Initial model parameters can be selected based on previous rithm has been proved using a Lyapunov function. The soft
learned model parameters to speed up tracking and leaning robotic glove was developed based on a SECA containing
convergence. an elastic plate as a torque compensating layer to overcome
In future work, rehabilitation training for stroke survi- finger spasticity. We have constructed a second-order linear
vors will be conducted to test the training performance of system for the actuator–finger system. Based on this
the designed controller and evaluate the rehabilitation effec- model, an online learning adaptive control method is able
tiveness after 20 sessions of training. The model parameters to provide the model parameters for the whole system. This
can record finger conditions during training sessions. This gives us an insight into the actuator–finger system and the
will help researchers to understand the behavior of spasti- finger alone. The proposed method has been evaluated
city after stroke, as well as to assess changes in spasticity with three settings: the actuator alone, the actuator with
during the rehabilitation training sessions. unimpaired fingers, and the actuator with fingers affected
by stroke. The results show that the stiffness term for the
actuator–stroke finger is significantly larger than that for
7.2. Model limitations the actuator–healthy finger. This implies that the proposed
Our proposed model learning adaptive control method is method could be developed further as a novel approach for
based on a linear model structure. The soft actuator–finger quantitative evaluation of finger condition in stroke
system consists of a SECA model and a finger joint model. survivors.
Funding and experimental results. In: 2012 IEEE International Confer-

The author(s) disclosed receipt of the following financial support ence on Robotics and Automation (ICRA). IEEE, pp. 279–284.
for the research, authorship, and/or publication of this article: This Braganza D, Dawson DM, Walker ID and Nath N (2007) A neural
work was supported by the CUHK T Stone Robotics Institute network controller for continuum robots. IEEE Transactions
(grant number C4930759), the Knowledge Transfer Project Fund on Robotics 23(6): 1270–1277.
(grant number KPF18HLF12) of the Chinese University of Hong Bukkems B, Kostic D, De Jager B and Steinbuch M (2005)
Kong, and the Innovation and Technology Fund of Hong Kong Learning-based identification and iterative learning control
(grant number ITS/065/18FP). of direct-drive robots. IEEE Transactions on Control Sys-
tems Technology 13(4): 537–549.
Catalano MG, Grioli G, Farnioli E, Serio A, Piazza C and Bicchi
ORCID iD A (2014) Adaptive synergies for the design and control of the
Zhi Qiang Tang https://orcid.org/0000-0001-6555-9938 pisa/iit softhand. The International Journal of Robotics
Research 33(5): 768–782.
Chen FC, Appendino S, Battezzato A, Favetto A, Mousavi M and
Supplemental material Pescarmona F (2014) Human finger kinematics and dynamics.
Supplemental material for this article is available online. In: Petuya V, Pinto C and Lovasz EC (eds.) New Advances in
Mechanisms, Transmissions and Applications. Dordrecht:
Springer, pp. 115–122.
References Cueli JR and Bordons C (2008) Iterative nonlinear model predic-
Amirabdollahian F, Loureiro R, Driessen B and Harwin W (2001) tive control. stability, robustness and applications. Control
Error correction movement for machine assisted stroke rehabi- Engineering Practice 16(9): 1023–1034.
litation. Integration of Assistive Technology in the Information Deimel R and Brock O (2016) A novel type of compliant and
Age 9: 60–65. underactuated robotic hand for dexterous grasping. The Inter-
Amirabdollahian F, Loureiro R and Harwin W (2002) Minimum national Journal of Robotics Research 35(1–3): 161–185.
jerk trajectory control for rehabilitation and haptic applications. In: Dempster WT (1955) Space requirements of the seated operator,
Proceedings IEEE International Conference on Robotics and geometrical, kinematic, and mechanical aspects of the body
Automation, 2002 (ICRA’02), Vol. 4. IEEE, pp. 3380–3385. with special reference to the limbs. Technical report, Michigan
Ansari NN, Naghdi S, Arab TK and Jalaie S (2008) The interrater State University East Lansing.
and intrarater reliability of the modified ashworth scale in the Dovat L, Lambercy O, Gassert R, et al. (2008) Handcare: A cable-
assessment of muscle spasticity: Limb and muscle group effect. actuated rehabilitation system to train hand function after
NeuroRehabilitation 23(3): 231–237. stroke. IEEE Transactions on Neural Systems and Rehabilita-
Ao D, Song R and Tong Ky (2015) Sensorimotor control of track- tion Engineering 16(6): 582–591.
ing movements at various speeds for stroke patients as well as Galloway KC, Becker KP, Phillips B, et al. (2016) Soft robotic
age-matched and young healthy subjects. PLoS One 10(6): grippers for biological sampling on deep reefs. Soft Robotics
e0128328. 3(1): 23–33.
Arimoto S, Kawamura S and Miyazaki F (1984) Bettering opera- George Thuruthel T, Ansari Y, Falotico E and Laschi C (2018)
tion of dynamic systems by learning: A new control theory for Control strategies for soft robotic manipulators: A survey. Soft
servomechanism or mechatronics systems. In: The 23rd IEEE Robotics 5(2): 149–163.
Conference on Decision and Control. IEEE, pp. 1064–1069. Gerboni G, Diodato A, Ciuti G, Cianchetti M and Menciassi A
Awad LN, Bae J, O’Donnell K, et al. (2017) A soft robotic exosuit (2017) Feedback control of soft robot actuators via commer-
improves walking in patients after stroke. Science Translational cial flex bend sensors. IEEE/ASME Transactions on Mechatro-
Medicine 9(400): eaai9084. nics 22(4): 1881–1888.
Balasubramanian S, Klein J and Burdet E (2010) Robot-assisted Hesse S, Kuhlmann H, Wilk J, Tomelleri C and Kirker SG (2008)
rehabilitation of hand function. Current Opinion in Neurology A new electromechanical trainer for sensorimotor rehabilitation of
23(6): 661–670. paralysed fingers: A case series in chronic and acute stroke
Balasubramanian S, Wei R, Perez M, et al. (2008) Rupert: An patients. Journal of Neuroengineering and Rehabilitation 5(1): 21.
exoskeleton robot for assisting rehabilitation of arm functions. Heung HL, Tong KY, Lau TH and Li Z (2019) Robotic glove with
In: Virtual Rehabilitation, 2008. IEEE, pp. 163–167. soft-elastic composite actuators for assisting activities of daily
Banerjee A, Bhattacharya B and Mallik A (2008) Large deflection living. Soft Robotics 6(2): 289–304.
of cantilever beams with geometric non-linearity: Analytical Hogan N (1984) An organizing principle for a class of voluntary
and numerical approaches. International Journal of Non-Lin- movements. Journal of Neuroscience 4(11): 2745–2754.
ear Mechanics 43(5): 366–376. Hu X, Tong K, Wei X, Rong W, Susanto E and Ho S (2013) The
Best CM, Gillespie MT, Hyatt P, Killpack M, Rupert L and Sher- effects of post-stroke upper-limb training with an electromyo-
rod V (2016) Model predictive control for pneumatically actu- graphy (EMG)-driven hand robot. Journal of Electromyogra-
ated soft robots. IEEE Robotics and Automation Magazine phy and Kinesiology 23(5): 1065–1074.
2(9): 31. Huang HW, Tibbitt MW, Huang TY and Nelson BJ (2018)
Blana D, van den Bogert AJ, Murray WM, et al. (2019) Model- Matryoshka-inspired micro-origami capsules to enhance load-
based control of individual finger movements for prosthetic ing, encapsulation, and transport of drugs. Soft Robotics, in
hand function. bioRxiv 629246. press.
Bouffard P, Aswani A and Tomlin C (2012) Learning-based model Ichikawa S, Tomita M, Doki S and Okuma S (2006) Sensorless
predictive control on a quadrotor: Onboard implementation control of permanent-magnet synchronous motors using online
Tang et al. 19
parameter identification based on system identification theory. Parks P (1966) Liapunov redesign of model reference adaptive
IEEE Transactions on Industrial Electronics 53(2): 363–372. control systems. IEEE Transactions on Automatic Control
In H, Kang BB, Sin M and Cho KJ (2015) Exo-glove: A wearable 11(3): 362–367.
robot for the hand with a soft tendon routing system. IEEE Peters M, Mackenzie K and Bryden P (2002) Finger length and
Robotics and Automation Magazine 22(1): 97–105. distal finger extent patterns in humans. American Journal of
Ioannou PA and Sun J (1996) Robust Adaptive Control, Vol. 1. Physical Anthropology: The Official Publication of the
Upper Saddle River, NJ: Prentice-Hall. American Association of Physical Anthropologists 117(3):
Kamper DG, Fischer HC, Cruz EG and Rymer WZ (2006) Weak- 209–217.
ness is the primary contributor to finger impairment in chronic Polygerinos P, Wang Z, Galloway KC, Wood RJ and Walsh CJ
stroke. Archives of Physical Medicine and Rehabilitation (2015) Soft robotic glove for combined assistance and at-home
87(9): 1262–1269. rehabilitation. Robotics and Autonomous Systems 73:
Kamper DG, Hornby TG and Rymer WZ (2002) Extrinsic flexor 135–143.
muscles generate concurrent flexion of all three finger joints. Rao SS and Yap FF (2011) Mechanical Vibrations, Vol. 4. Upper
Journal of Biomechanics 35(12): 1581–1589. Saddle River, NJ: Prentice Hall.
Kim D, Kwon J, Han S, Park YL and Jo S (2018) Deep full-body Reinhart RF, Shareef Z and Steil JJ (2017) Hybrid analytical and
motion network for a soft wearable motion sensing suit. IEEE/ data-driven modeling for feed-forward robot control. Sensors
ASME Transactions on Mechatronics 24(1): 56–66. 17(2): 311.
Kocijan J (2016) Modelling and Control of Dynamic Systems Renda F, Giorgio-Serchi F, Boyer F, Laschi C, Dias J and Senevir-
using Gaussian Process Models. Berlin: Springer. atne L (2018) A unified multi-soft-body dynamic model for
Lee K (2002) Large deflections of cantilever beams of non-linear underwater soft robots. The International Journal of Robotics
elastic material under a combined loading. International Jour- Research 37(6): 648–666.
nal of Non-Linear Mechanics 37(3): 439–443. Soter G, Conn A, Hauser H and Rossiter J (2018) Bodily aware
Lee KS, Chin IS, Lee HJ and Lee JH (1999) Model predictive soft robots: Integration of proprioceptive and exteroceptive
control technique combined with iterative learning for batch sensors. In: 2018 IEEE International Conference on Robotics
processes. AIChE Journal 45(10): 2175–2187. and Automation (ICRA). IEEE, pp. 2448–2453.
Leibovic SJ and Bowers WH (1994) Anatomy of the proximal Susanto EA, Tong RK, Ockenfeld C and Ho NS (2015) Efficacy
interphalangeal joint. Hand Clinics 10(2): 169–178. of robot-assisted fingers training in chronic stroke survivors: A
Li Z and Du R (2013) Design and analysis of a bio-inspired wire- pilot randomized-controlled trial. Journal of Neuroengineering
driven multi-section flexible robot. International Journal of and Rehabilitation 12(1): 42.
Advanced Robotic Systems 10(4): 209. Takahashi CD, Der-Yeghiaian L, Le V, Motiwala RR and Cramer
Liu X and Xi K (2014) Feedback-assisted iterative learning model SC (2007) Robot-based hand motor therapy after stroke. Brain
predictive control with nonlinear fuzzy model. Mathematical 131(2): 425–437.
Problems in Engineering 2014: 874705. Tang ZQ, Heung HL, Tong KY and Li Z (2019) A novel iterative
Loureiro R, Amirabdollahian F, Coote S, Stokes E and Harwin W learning model predictive control method for soft bending
(2001a) Using haptics technology to deliver motivational thera- actuators. In: 2019 IEEE International Conference on Robotics
pies in stroke patients: Concepts and initial pilot studies. In: and Automation (ICRA). IEEE, pp. 4004–4010.
Proceedings of EuroHaptics, p. 6. Thibaut A, Chatelle C, Ziegler E, Bruno MA, Laureys S and Gos-
Loureiro R, Amirabdollahian F, Driessen B and Harwin W series O (2013) Spasticity after stroke: Physiology, assessment
(2001b) A novel method for computing natural path for and treatment. Brain Injury 27(10): 1093–1105.
robot assisted movements in synthetic worlds. In: Marincek C, Thuruthel TG, Shih B, Laschi C and Tolley MT (2019) Soft robot
Buhler C, Knops H and Andrich R (eds) Rehabilitation, Allied perception using embedded soft sensors and recurrent neural
Medicine: Assistive Technology - Added Value to the Quality networks. Science Robotics 4(26): eaav1488.
of Life (AAATE’01) (Assistive Technology Research Series, Timmermans AA, Seelen HA, Willmann RD and Kingma H
No. 10). Amsterdam: IOS Press, pp. 262–267. (2009) Technology-assisted training of arm–hand skills in
Marchese AD and Rus D (2016) Design, kinematics, and control stroke: Concepts on reacquisition of motor control and thera-
of a soft spatial fluidic elastomer manipulator. The Interna- pist guidelines for rehabilitation technology design. Journal of
tional Journal of Robotics Research 35(7): 840–869. Neuroengineering and Rehabilitation 6(1): 1.
Marchese AD, Tedrake R and Rus D (2016) Dynamics and trajec- Tong K, Ho S, Pang P, et al. (2010) An intention driven hand func-
tory optimization for a soft spatial fluidic elastomer manipula- tions task training robotic system. In: 2010 Annual Interna-
tor. The International Journal of Robotics Research 35(8):
tional Conference of the IEEE Engineering in Medicine and
1000–1019.
Biology. IEEE, pp. 3406–3409.
Mascarenhas WF (2014) The divergence of the BFGS and Gauss
Trimmer B (2014) Soft robot control systems: A new grand chal-
Newton methods. Mathematical Programming 147(1-2):
lenge? Soft Robotics 1(4): 231–232.
253–276.
Underwood SJ and Husain I (2010) Online parameter estimation
Morrow J, Shin HS, Phillips-Grafflin C, et al. (2016) Improving
and adaptive control of permanent-magnet synchronous
soft pneumatic actuator fingers through integration of soft sen-
machines. IEEE Transactions on Industrial Electronics 57(7):
sors, position and force control, and rigid fingernails. In: ICRA,
2435–2443.
pp. 5024–5031.
Villoslada Á, Rivera C, Escudero N, Martı́n F, Blanco D and Mor-
Park YL, Chen Br, Pérez-Arancibia NO, et al. (2014) Design and
eno L (2018) Hand exo-muscular system for assisting
control of a bio-inspired soft wearable robotic device for
astronauts during extravehicular activities. Soft Robotics, in
ankle–foot rehabilitation. Bioinspiration and Biomimetics 9(1):
press.
016007.
Webster III RJ and Jones BA (2010) Design and kinematic model- ~i

Ei = qi ^qi = Fi u Fi ui = Fi u
ing of constant curvature continuum robots: A review. The ð28Þ
Dui = ui + 1 ui = r(FTi Fi )y FTi Ei = ru
~i
International Journal of Robotics Research 29(13):
1661–1683.
Yap HK, Khin PM, Koh TH, et al. (2017) A fully fabric-based Then, we have
bidirectional soft robotic glove for assistance and rehabilitation
of hand impaired patients. IEEE Robotics and Automation Let- ~T r1 u
Fi + 1 Fi = u ~i + 1 u ~i
~T r1 u
i+1 i
ters 2(3): 1383–1390. ~i + Dui )T r1 Dui
= (2u
Yap HK, Lim JH, Goh JCH and Yeow CH (2016) Design of a soft ð29Þ
robotic glove for hand rehabilitation of stroke patients with = (r 2)u ~
~T u
i i
clenched fist deformity using inflatable plastic actuators. Jour- 2
nal of Medical Devices 10(4): 044504. = (r 2)u~i
2
You X, Zhang Y, Chen X, et al. (2017) Model-free control for soft
~i are
Because r\2, Fi is non-increasing. Then Fi , ui and u
manipulators based on reinforcement learning. In: 2017 IEEE/
RSJ International Conference on Intelligent Robots and Sys- bounded. In addition, we also have
tems (IROS). IEEE, pp. 2909–2915. 2
Yue Z, Zhang X and Wang J (2017) Hand rehabilitation robotics Fi + 1 Fi = r1 u ~i + 1 2 u~i \0
2 2
on poststroke motor recovery. Behavioural Neurology 2017: 2 2
) u ~i + 1 \u ~i ð30Þ
3908135. 2 2
Zhang Z, Dequidt J, Kruszewski A, Largilliere F and Duriez C ) kui + 1 u k22 \kui u k22
(2016) Kinematic modeling and observer based control of soft
robot using real-time finite element method. In: 2016 IEEE/ Inequality (30) implies that the error between estimated
RSJ International Conference on Intelligent Robots and Sys- parameter value and true model parameter value is monoto-
tems (IROS). IEEE, pp. 5509–5514. nically decreasing. Moreover, when r 2 (0, 1), we have
Fi + 1 \(1 r)Fi
Appendix A. Index to multimedia extensions ) Fi + n \(1 r)n Fi ð31Þ
Archives of IJRR multimedia extensions published prior to ) Fi \(1 r)i F0
2014 can be found at http://www.ijrr.org, after 2014 all 2
As Fi = r1 u
~i , we have
videos are available on the IJRR YouTube channel at http:// 2
www.youtube.com/user/ijrrmultimedia 2
r1 u
~i \(1 r)n Fi
22 ð32Þ
Table of Multimedia Extensions ) u ~i \F0 × r(1 r)i
2
2
Extension Media type Description and as F0 = r1 u
~0 = r1 ku0 u k2 , we have
2 2
1 Video Free bending situation. 2

2 Video Healthy subject situation. u
~i \u0 u × (1 r)i ð33Þ
2
3 Video Stroke subject situation. 2
Inequality (33) shows that u ~i approaches zero asympto-
2
tically as i goes to infinity when r 2 (0, 1). This implies
that ui asymptotically converges to u . In addition, we can
Appendix B: Convergence analysis see that a proper initial value that is close to true value will
help with faster convergence. The convergence speed will
Theorem 1. If r 2 (0, 1), then ui asymptotically converges also be faster when r is given a larger value in the range
to the true model parameter u based on Equations (14)– (0, 1). h
(18).
Theorem 2. if l satisfies 0\(1 lG z)\1 where G rep-
resents the true SECA–finger system model, then the track-
Proof. We first define the parameter change as ing error asymptotically converges to zero based on
~i = ui u . Then we choose the Lyapunov function
u Equations (19)–(22).
(Ioannou and Sun, 1996) to evaluate the rate of parameter
change, shown as follows:
Proof. Here we use G to represent the true SECA–finger
system model that we want to find after model learning
~T r1 u
Fi = u ~i ð27Þ
i process:
From Equations (17) and (18), we know that
yi (k) = G Pi (k) ð34Þ
Tang et al. 21
Based on Equations (20), (21), and (22), we can analyze ei + 1 (k + 1) ł ei (k + 1) G lei (k + 2) ð36Þ
the reference tracking error as follows: We take z-transformation on both sides of inequality
(36), then we have
ei + 1 (k + 1) = r(k + 1) yi + 1 (k + 1)
= ei (k + 1) + yi (k + 1) yi + 1 (k + 1) zEi + 1 (z) ł zEi (z) z2 G lEi (z)
= ei (k + 1) + G Pi (k + 1) G Pi + 1 (k + 1) ) Ei + 1 (z) ł (1 lG z)Ei (z)
k +1 k +1
= ei (k + 1) G lei (k + 2) + G (hi + 1êi + 1 hiêi ) ð37Þ
) Ei + k (z) ł (1 lG z)k Ei (z)
ð35Þ ) Ei (z) ł (1 lG z)i E0 (z)
Inequality (37) shows that when we choose proper l such
As model parameters approach the true value with the
that 0\(1 lG z)\1, then the tracking error asymptoti-
increase of i, so the model accuracy increases with i. This
cally converges to zero as i approaches infinity. h
means the error between model predicted angle and true
angle decreases with i. It is reasonable to assume that
k +1 k +1 Appendix C: Nomenclature
hi + 1êi + 1 ł hiêi . Then we have
j SECA segment or finger joint: j = m for MCP; j = p for PIP

P, qj , q_ j , €qj Pressure, angle, angle velocity, and angle acceleration of joint j, respectively
r1 , r2 , w, d, t Inner radius, outer radius, chamber gap, thickness of SECA and thickness of torque compensating layer,
respectively
Vj , lj Material volume and length of SECA segment j, respectively
Es , Et Young’s modulus of the soft elastomer body and compensating layer, respectively
Kjs , Kjt Stiffness of the soft elastomer body and torque compensating layer at segment j, respectively
PEjs , PEjt Strain energy of the soft elastomer body and torque compensating layer at segment j, respectively
msj , mtj Mass of the soft elastomer body and the torque compensating layer at segment j, respectively
Jja , Daj , Kja , Tja Moment inertia, damping coefficient, stiffness, and moment–pressure constant of SECA segment j,
respectively
Mja , Mjf Moment on segment j of the SECA and joint j of finger, respectively
Jjf , Dfj , Kjf , Tjf Moment inertia, damping coefficient, stiffness, and moment–pressure constant of finger joint j, respectively
Kjf =Jjf , Dfj =Jjf , Tjf =Jjf Stiffness indicator, damping indicator, and moment–pressure indicator of finger joint j, respectively
A, B, C Coefficient matrix/vector of state space equation
y Total finger angle (the summation of MCP angle and PIP angle)
J j , Dj , K j , T j Moment inertia, damping coefficient, stiffness, and moment–pressure constant of SECA–finger system joint j,
respectively
r Reference trajectory
a0 , a1 , a2 , a3 , a4 , a5 Polynomial coefficients of reference trajectory
T , Ts Time period of reference trajectory and sampling time, respectively
r0 , re Start and end angles of reference trajectory, respectively
i, k Iteration number and time instance, respectively
^q, u, F Model estimated angle, estimated model parameter, and model information matrix, respectively
r, l, m Learning step size, feedforward input gain, and initial model parameter, respectively
Q, R, H Output, input weighting factors, and prediction horizon of MPC cost function, respectively
u, v Feedback and feedforward pressure, respectively
eri , epi Reference tracking error and model learning error at iteration i, respectively
dE Error tolerance

Model-Based Online Learning and Adaptive Control For A Human-Wearable Soft Robot'' Integrated System

Uploaded by

Copyright:

Available Formats

Model-Based Online Learning and Adaptive Control For A Human-Wearable Soft Robot'' Integrated System

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Model-Based Online Learning and Adaptive Control For A Human-Wearable Soft Robot'' Integrated System

Uploaded by

Copyright:

Available Formats

Article

The International Journal of

control for a ‘‘human-wearable soft robot’’ Article reuse guidelines:

1. Introduction upper (Balasubramanian et al., 2008) and lower (Park

Fig. 1. 3D-printed soft robotic glove.

the SECA to minimize any gap between the finger and

3. System modeling manipulator can be expressed as a function of bending

3.3. SECA–finger model 4.1. Reference angle design

r(t) = a0 + a1 t + a2 t2 + a3 t3 + a4 t4 + a5 t5 ð11Þ minkEi k22

d2r d2r where the learning step size r is a constant.

Table 1. Simulation control parameters.

m1 08 908 6s 0.1 s 100 1 20 0.4 10 1 × 1025

x_ = Ax + BP angle tracking can be achieved with satisfactory

The model described by (25) is assumed to be unavail-

Table 2. Simulation model parameters.

Km =Jm Dm =Jm Tm =Jm Kp =Jp Dp =Jp Tp =Jp

True model 300 5 3 400 10 5

(a) (b) (c)

(a) (b) (c)

(d) (e) (f)

Table 3. Stroke characteristics.

1 55–59 21 months Male Ischemic Left 9 1

Table 4. Experimental control parameters.

Free bending m1 108 808 6s 0.1 s 100 1 20 0.02 10 0.01

Table 5. Experimental learned SECA–finger system model parameters.

Km =Jm Dm =Jm Tm =Jm Kp =Jp Dp =Jp Tp =Jp

Free bending 51.21 5.10 0.25 5.11 3.74 0.03

6.1. Free bending case 6.2. Healthy subject case

From Figure 12 we can see that it takes 17 iterations to 7. Discussion

Table 6. Finger joint model parameters.

Healthy subject 1 100.59 9.20 0.49 72.11 4.72 0.59

Funding and experimental results. In: 2012 IEEE International Confer-

Webster III RJ and Jones BA (2010) Design and kinematic model- ~i

1 Video Free bending situation. 2

j SECA segment or finger joint: j = m for MCP; j = p for PIP

You might also like